The Circular Dollar: Why Part of AI's Boom Is the Industry Lending Money to Its Own Customers

Part of AI's reported growth is the industry's own money, booked twice. The mechanism is vendor financing: a supplier lends, guarantees, or invests the money its customers use to buy the supplier's product, so one dollar leaves the supplier's balance sheet, becomes the customer's purchasing power, and returns as the supplier's revenue. Nvidia has committed up to $105 billion in credit and compute to an OpenAI data center built to run Nvidia systems. SpaceX is paying $60 billion for an AI coding startup while booking a rival's compute rent as its own AI revenue under a contract worth roughly $15 billion a year. The growth is real on paper, and part of it is self-supplied.

It is the bar that hands patrons cash and counts the drinks: sales boom, the taps run, and the tab sits on the house's own card. Nothing is fake. The drinks get poured. The money is the bar's.

How one dollar books growth on both sides of the wall

The loop has four steps. The supplier extends credit, a guarantee, or an equity stake to the customer, or to the customer's landlord, reseller, or power supply. The customer uses that support to buy the supplier's product. The supplier books the sale as revenue. The support sits on the supplier's balance sheet as a receivable or a stake that only tells the truth in a downturn. The customer reports growth, the supplier reports growth, and the money never leaves the industry's books.

Vendor financing is legal, usually disclosed, and sometimes sensible: these chips do generate revenue for their buyers, and a supplier with a fortress balance sheet can rationally prime demand it believes in. The accounting is not the scandal. The structural problem is concentration. Once suppliers fund the buying, headline growth stops measuring outside demand and starts partly measuring supplier confidence. And supplier confidence is pro-cyclical: it peaks exactly when the risk peaks, which is why the loop that inflates a boom also concentrates its bust.

The deals that close the loop, and the tell each one carries

• Nvidia and OpenAI, Ohio. The two discussed a backstop of up to $250 billion that would let OpenAI raise debt for a 10-gigawatt campus 1. What a securities filing showed was up to $105 billion in credit and compute, about 42 cents on every dollar discussed 1. The circle closes three ways inside that filing: Nvidia finances the campus, Nvidia invests $1.5 billion in SB Energy, the company building it under a 20-year lease to OpenAI, and OpenAI itself holds a stake in SB Energy 1. The chipmaker is creditor to its customer, investor in its customer's landlord, and vendor into the campus.
• Nvidia and the bottlenecks. The chipmaker took a minority stake in Cloverleaf Infrastructure, a company that lays the groundwork for data centers, with the investment likely worth several hundred million dollars 2. It reportedly agreed to buy the AI model repository Hugging Face for $12.9 billion, a company founded in 2016 that is reportedly not yet profitable 3. And it stood up financing platforms with six large asset managers to deploy $500 billion of third-party capital into the buildout, signed as memorandums of understanding 1; 4. Nvidia's framing is that chips have become "an investable asset class" 4; CNBC's framing of the same maneuvers is that they "raised concerns about circular financing in the AI trade" 1.
• Stripe and OpenRouter. In August 2026, days after Stripe reportedly finalized a deal to acquire the AI gateway startup OpenRouter for more than $7 billion 5, OpenRouter's board lists OpenAI's flagship GPT-5.6 Sol at 50 percent off across OpenAI's own hosting tiers, while the Amazon Bedrock and Azure listings for the same model carry no discount 6. OpenAI's own price announcement cut two cheaper models, Terra by 20 percent and Luna by 80 percent, and left the flagship alone: "Sol's pricing remains the same" 7. On the price sheets in view, the only half-price Sol listing sits at a reseller, below the vendor's own list.
• SpaceX and Cursor. SpaceX closed its acquisition of the AI coding startup Cursor at the $60 billion option price set in its April deal 8. Anthropic, a rival the deal is meant to help catch, agreed to pay SpaceX close to $45 billion over three years for compute, roughly $15 billion a year 9. At full run, one rival's rent retires about a quarter of the Cursor purchase price every year. Elon Musk told staff AI revenue would exceed all other SpaceX revenue, probably by September 10; in the second quarter, SpaceX's AI segment brought in $2.56 billion against $5.25 billion for everything else, so the claim requires the segment to more than double in a single quarter 10. Anthropic's rent at full run, about $3.75 billion a quarter, is larger than the roughly $2.7 billion gap.

Four tells that classify the next announcement yourself

1. Guarantee or financing language attached to a purchase commitment. The Ohio backstop is the current example; Broadcom's $35 billion package with Apollo and Blackstone, with about a million chips posted as collateral and a Broadcom guarantee on the senior notes, is the same tell from a different supplier 4. Ask: if the buyer cannot pay, who repays, and does the guarantor also sell the product?
2. Supplier equity flowing into a customer, a reseller, or a bottleneck. A supplier buying into the grid, the model commons, or its customer's landlord is buying the constraints its customers live under 2; 3. Ask: whose balance sheet now controls your bottleneck?
3. A price cut below the vendor's list. When a reseller sells below the vendor's list price and the vendor holds its own price, the discount is funded inside the chain, not by the vendor's price sheet 6; 7. Ask: whose margin thins when volume slows?
4. Revenue recognized from a counterparty the recognizer funds, equips, or acquires. SpaceX's AI segment counting rivals' rent, and Nvidia booking sales into campuses it finances, are the same shape 9; 1. Ask: what share of this "AI revenue" comes from counterparties the seller also bankrolls?

Telecom already ran this experiment

In 1999 Lucent was the sixth-largest company in America by market capitalization, and Nortel alone accounted for more than a third of the Toronto Stock Exchange's value 11. Equipment makers from Lucent and Nortel to Motorola, Nokia, and Cisco extended loans to telecom startups "often after those companies were refused loans by real banks," financing not just the gear but the customers' businesses 12. Peak vendor lending hit $25 to $30 billion in 2000, with financing offers reportedly reaching as high as 200 percent of the equipment's value during the 3G license auctions 12.

Scale is what separates now from then. Nvidia's single Ohio commitment, up to $105 billion, is roughly four times the entire equipment industry's lending at its 2000 peak 1; 12. The break, when it came, was fast. Lucent's bad loans went from 2.6 percent of total loans at the end of 2000 to 60 percent at the end of 2001, a twenty-three-fold jump in one reporting year, and Nortel's went from 25.5 percent to 80 percent 12. Eighteen of the 37 publicly listed US telecom service providers founded since 1996 had gone bankrupt by late 2002 12. Lucent ended up a sliver of itself, sold to Alcatel, and Nortel went bankrupt 11. Vendors kept lending past the top because their own stock prices were tied to the fate of the customers they had created 12. One adviser's summary of that extra year of lending: "It bought them one more year" 12.

What breaks first when the circle stops

• Guarantees walk back first. Ohio's $250 billion discussion became a $105 billion filing 1. A walk-back is not the loop ending; it is the supplier trimming its own tail risk while keeping demand lit, which is what telecom vendors did after March 2000 12.
• Exit clauses convert revenue into vacancy. The Anthropic agreement unwinds on 90 days' notice, at a company that spent over $25 billion on AI in one quarter 10.
• Receivables poison within a year. Lucent's one-year jump from 2.6 to 60 percent bad loans is the template 12. The AI-era version will surface inside the new financing platforms, $500 billion of third-party capital in what The Verge calls GPU-backed lending, against chips as the asset class 4. As supplier risk moves onto those platforms, the live price of it becomes the yield they pay to borrow, and that yield is the number to watch as these deals mature.
• Subsidized discounts reverse. A reseller selling at 50 percent off runs the meter on margin somewhere inside the chain 6. If volume does not grow into it, the discount is the first line to disappear, and the demand it manufactured goes with it.

The one-sentence test

For any AI announcement, ask: whose money is buying, and does any of it return to the seller? If the buyer's funding traces back to the seller's guarantee, equity, discount, or credit, you are looking at the Circular Dollar: growth that partly never leaves the room. That does not make the deals fake. The chips are real, the rent is contracted, the accounting is legitimate. It makes them correlated. When an industry lends money to its own customers, the boom borrows its demand from a handful of supplier balance sheets, and those balance sheets decide, quarter by quarter, how much of the boom keeps existing. Two honest unknowns remain: how much of the headline growth is supplier-funded in aggregate, which no disclosure we reviewed totals, and whether the memorandums behind the biggest platforms harden into deals, given that Nvidia's earlier $100 billion OpenAI memorandum never became one 4. Telecom's lenders learned in 2001 what that correlation costs. The four tells are how you watch for the repeat, one announcement at a time.